Neural network learning engine

Abstract

A neural network for assertion-based scoring can include a support vector classifier and a plurality of sequential layers. The sequential layers can include, for example, a word embedding layer, a conventional layer, a recurrent layer, a dense layer, and/or a dropout layer. The neural network can be configured to perform a word-frequency count and/or can be configured to collect candidate words of various Levenshtein distances from standard-spelling variations.

I claim: 
     
         1 . A neural network for assertion-based scoring, comprising:
 a support vector classifier; and   a plurality of sequential layers, wherein the plurality of sequential layers include a word embedding layer, a conventional layer, a recurrent layer, a dense layer, and a dropout layer.   
     
     
         2 . The neural network of  claim 1 , wherein the word embedding layer is configured to perform word embedding. 
     
     
         3 . The neural network of  claim 2 , wherein word embedding is a single component of a larger neural network. 
     
     
         4 . A neural network for assertion-based scoring, comprising:
 a support vector classifier;   a plurality of sequential layers, wherein the plurality of sequential layers include a word embedding layer, a first convolutional layer, a second convolutional layer, a recurrent layer, and a plurality of dense layers; and   wherein the recurrent layer forms a component of the scoring engine.   
     
     
         5 . The neural network of  claim 4 , wherein the recurrent layer utilizes gated-recurrent-units. 
     
     
         6 . The neural network of  claim 4 , wherein the recurrent layer utilizes a simple recurrent neural network. 
     
     
         7 . The neural network of  claim 4 , wherein the recurrent layer utilizes a long-short-term-memory network. 
     
     
         8 . The neural network of  claim 4 , wherein the recurrent layer utilizes a neural architecture search. 
     
     
         9 . The neural network of  claim 4 , wherein the plurality of dense layers utilizes filters. 
     
     
         10 . The neural network of  claim 9 , wherein the filters are of size 256, 64, and 16. 
     
     
         11 . A neural network for assertion-based scoring, comprising:
 a support vector classifier; and   a plurality of sequential layers, wherein the neural network is configured to perform scoring, and wherein the scoring is short answer scoring.   
     
     
         12 . A neural network for assertion-based scoring, comprising:
 a support vector classifier;   a plurality of sequential layers, wherein the neural network is configured to perform a word count on a text, wherein each word has a frequency in the text; and   wherein the neural network is configured to store the frequency of each word.   
     
     
         13 . The neural network of  claim 12 , wherein the frequency of each word is normalized to generate a normalized frequency of each word. 
     
     
         14 . The neural network of  claim 12 , wherein the neural network is further configured to calculate a weight, w, for each word according to the equation:
     w =log(( i+ 1)log(length of the word)),   where the integer i is the order in the vocabulary, which is ordered by listing the most frequent words first.   
     
     
         15 . A neural network for assertion-based scoring, comprising:
 a support vector classifier;   a plurality of sequential layers, wherein the neural network is configured to determine whether a word in a training set is included in an embedding;   wherein the neural network is configured to identify misspelled words from the text; and wherein the neural network further comprises a spell-correction engine and a word embedding, wherein the word embedding includes standard-spelling variations.   
     
     
         16 . The neural network of  claim 15 , wherein the neural network is configured to collect a first set of candidates, wherein the first set includes all words from the text that are of Levenshtein distance 1 from the standard-spelling variations. 
     
     
         17 . The neural network of  claim 16 , wherein if the first set of candidates is not empty, the neural network is configured to return the word having the highest frequency in the text. 
     
     
         18 . The neural network of  claim 17 , wherein if the first set of candidates is empty, the neural network is configured to collect a second set of candidates, wherein the second set includes all words from the text that are of Levenshtein distance 2 from the standard-spelling variations. 
     
     
         19 . The neural network of  claim 18 , wherein if the second set of candidates is not empty, the neural network is configured to return the word having the highest frequency in the text. 
     
     
         20 . The neural network of  claim 19 , wherein if the second set of candidates is empty, the neural network is configured to partition inputs into words that minimize the sum of the word weights. 
     
     
         21 . The neural network of  claim 15 , wherein the word is a Levenshtein distance of 3 from a standard-spelling variation, and the neural network is configured to search the word of Levenshtein distance of 3. 
     
     
         22 . The neural network of  claim 15 , wherein the word is a Levenshtein distance of 4 from a standard-spelling variation, and the neural network is configured to search the word of Levenshtein distance of 4. 
     
     
         23 . The neural network of  claim 15 , wherein the word is a Levenshtein distance of 5 from a standard-spelling variation, and the neural network is configured to search the word of Levenshtein distance of 5. 
     
     
         24 . The neural network of  claim 15 , wherein the word is a Levenshtein distance of between 1 and 5 from a standard-spelling variation, and the neural network is configured to search the word of Levenshtein distance of between 1 and 5.